# K-01 — LLM + Harness anatomy
# Category: KEY
# Kapitel: 02 – Anatomie des Coding-Agents
# Anker: Agent = LLM + Harness
# Orientierung: 16:9 landscape
# Didaktik: Zwei Schichten — Modell generiert Token; Harness liefert Tools, Loop, Permissions, Memory.
# Caption (DE): Agent = LLM + Harness — das Modell generiert, der Harness führt aus und speist Observations zurück.
# Ziel-Asset: assets/illustrations/K-01-llm-plus-harness.jpg
# Language: English (Gemini / Imagen)
# Usage: paste entire prompt body below into Gemini

--- PROMPT START ---
Create a clean educational technical illustration, 16:9 landscape, for a professional practice book about LLM-based coding agents.

TOPIC: A coding agent is LLM + Harness — two layers, one loop. Not a magical black-box “AI coworker”.

GLOBAL STYLE:
- flat vector + soft depth; deep navy (#0B1F33), teal (#0D9488), soft blue (#2563EB), off-white panels (#F8FAFC), slate (#334155)
- minimal English labels (2–6 words); no brand logos; no robot mascots; print-ready; high grayscale contrast

QUALITY GUARDRAILS:
- every visible word correctly spelled English; max ~12 short labels
- no prompt meta-text as labels; no vendor logos
- hero terms: LLM, Harness, Context, Tools

Title text exactly: "Agent = LLM + Harness"

COMPOSITION — two stacked layers + bottom loop ribbon:

1) TOP LAYER (soft blue panel, navy border), label: "LLM — token generator"
   Three small boxes inside only:
   - "read context"
   - "predict tokens"
   - "text / tool-calls"
   Tiny caption under layer: "no commit semantics"

2) BOTTOM LAYER (teal-tinted panel, navy border), label: "Harness — control runtime"
   Eight small tiles in two rows (short labels only):
   - "system prompt"
   - "tool interface"
   - "tool runtime"
   - "loop controller"
   - "permissions"
   - "context assembler"
   - "memory / AGENTS.md"
   - "UI: diffs / approve"

3) RIGHT SIDE of harness: two badges stacked
   - green: "read-ish: add context"
   - amber-red: "write-ish: change world"

4) BOTTOM FLOW RIBBON (full width):
   curved loop arrows with short chips:
   "Context → LLM decides → Tool/Text → Observation → Context"

MOOD: precise systems diagram for engineers; calm textbook quality.
ASPECT: 16:9 landscape. Generous margins for print.
NEGATIVE: no robots, no neon cyberpunk, no brand logos, no photoreal UI, no watermarks, no walls of code, max ~12 labels.
--- PROMPT END ---

# Follow-up edit:
# Keep two clear layers LLM vs Harness. Bottom loop must include Observation → Context.
# Preserve read-ish vs write-ish badges. Reduce clutter if labels exceed ~12.
